Back to results
US 7,689,520 B2Granted
Machine learning system and method for ranking sets of data using a pairing cost function
AssigneeMicrosoft Corporation
InventorsChristopher J. C. Burges, Tal Shaked
Life of this patent
Estimating the term…
Abstract
A machine learning system to rank data within sets is disclosed. The system comprises a ranking module that has differentiable parameters. The system further comprises a cost calculation module that uses a cost function that depends on pairs of examples and which describes an output of the ranking module. Methods of using the disclosed system are also provided.
Claims12
A document analysis machine learning system using a neural network model to rank data within sets, comprising:
- at least one processor;
- a ranking module coupled to the at least one processor and having differentiable parameters, each of the differentiable parameters having a plurality of potential weights, to determine a rank value for an input example based on the weights of differentiable parameters, wherein weights are trained with a training data set that includes pairs of ranked examples;
- and
- a cost calculation module coupled to the at least one processor that uses a ranking cost function to determine the weights associated with said differentiable parameters,
- wherein the cost function is an asymmetric and differentiable function that maps a result of a comparison of a target rank value of a pair of input examples with measured rank values of the pair of input examples to a real number,
- wherein the target rank value varies for each pair of input examples, and
- wherein an input example is represented by said differential parameters and the rank value of the input example is generated according to the weights associated with differential variables,
- wherein the weights are set according to a minimization of the total cost, which comprises the sum of plurality of cost associated with each pair of input examples,
- wherein the cost of each pair of input example weights are updated according to a difference of two terms, with a first term depending on the first example and a second term depending on the second example, multiplied by the derivative of the cost function;
- and
- a final output of the ranking system including a final weight assigned to the differentiable parameters used for ranking documents, wherein ordering a data set based upon the rank value is consistent with a final weight assigned for each of the differentiable parameters selected as a function of the cost function.
8 more elements
A computer-implemented document analysis using a neural network model method of ranking data, comprising:
- storing, in a memory, instructions for performing the method of ranking data using a neural network model;
- executing the instructions on a processor;
- according to the instructions being executed:
- determining a rank value for an input example utilizing a ranking module coupled to the processor,
- wherein the ranking module comprises differentiable parameters, each of the differentiable parameters having a plurality of potential weights,
- wherein the ranking module determines the rank value for an input example based on the weights of the differentiable parameters,
- wherein weights are trained with a training data set that includes pairs of ranked examples;
- and
- utilizing a cost calculation module coupled to the processor,
- wherein the cost calculation module uses a ranking cost function to determine the weights associated with said differentiable parameters,
- wherein the cost function is an asymmetric and differentiable function that maps a result of a comparison of a target rank value of a pair of input examples with measured rank values of the pair of input examples to a real number,
- wherein the target rank value varies for each pair of input examples, and wherein an input example is represented by said differential parameters and the rank value of the input example is generated according to the weights associated with differential variables,
- wherein the weights are set according to a minimization of the total cost, which comprises the sum of plurality of cost associated with each pair of input examples,
- wherein the cost of each pair of input example weights are updated according to a difference of two terms, with a first term depending on the first example and a second term depending on the second example, multiplied by the derivative of the cost function;
- and
- outputting a final output, wherein the final output of the ranking system includes a final weight assigned to the differentiable parameters used for ranking documents, wherein ordering a data set based upon the rank value is consistent with a final weight assigned for each of the differentiable parameters selected as a function of the cost function.
13 more elements
A computer readable medium encoded with instructions for ranking data utilizing a neural network model, the instructions implementing a method comprising:
- storing, in a memory, instructions for performing the method of ranking data using a neural network model;
- executing the instructions on a processor;
- according to the instructions being executed:
- determining a rank value for an input example utilizing a ranking module coupled to the processor,
- wherein the ranking module comprises differentiable parameters, each of the differentiable parameters having a plurality of potential weights,
- wherein the ranking module determines the rank value for an input example based on the weights of the differentiable parameters,
- wherein weights are trained with a training data set that includes pairs of ranked examples;
- and
- utilizing a cost calculation module coupled to the processor,
- wherein the cost calculation module uses a ranking cost function to determine the weights associated with said differentiable parameters,
- wherein the cost function is an asymmetric and differentiable function that maps a result of a comparison of a target rank value of a pair of input examples with measured rank values of the pair of input examples to a real number,
- wherein the target rank value varies for each pair of input examples, and wherein an input example is represented by said differential parameters and the rank value of the input example is generated according to the weights associated with differential variables,
- wherein the weights are set according to a minimization of the total cost, which comprises the sum of plurality of cost associated with each pair of input examples,
- wherein the cost of each pair of input example weights are updated according to a difference of two terms, with a first term depending on the first example and a second term depending on the second example, multiplied by the derivative of the cost function;
- and
- outputting a final output, wherein the final output of the ranking system includes a final weight assigned to the differentiable parameters used for ranking documents, wherein ordering a data set based upon the rank value is consistent with a final weight assigned for each of the differentiable parameters selected as a function of the cost function.
13 more elements
Dependencies are inferred from the claim text.